Synthetic Identity Fraud and Deepfakes: The New Front Line of U.S. Onboarding

When the Document Is Perfect and the Person Does Not Exist

American identity verification was built on two assumptions that no longer hold.

The first was that producing a convincing forged identity document required skill, equipment, and time. For most of the last century that was true. Forgery was a craft, and the population capable of it was small enough that document security features provided meaningful protection.

The second was that a Social Security Number, combined with a name and date of birth, reliably identified a specific human being. This assumption underpinned nearly every U.S. onboarding process, from bank account opening to investor subscription to tenant screening.

Generative AI dismantled the first assumption. Synthetic identity fraud dismantled the second. Together they produced a situation FinCEN documented directly in its November 2024 alert on deepfake media, which described a marked rise in suspicious activity reports involving GenAI-created identity documents used specifically to circumvent identity verification and authentication controls.

For U.S. funds, real estate firms, and professional services firms conducting KYC, this is not a distant banking problem. Any organization that accepts a photograph of an identity document as proof of identity is running a control that a motivated adversary can now defeat cheaply.

---

Two Distinct Threats, Frequently Confused

Practitioners often use "synthetic identity fraud" and "deepfake fraud" interchangeably. They are different problems requiring different controls.

Synthetic identity fraud

A synthetic identity combines real and fabricated data elements to create a person who does not exist but whose identity nonetheless passes verification. Typically this means a genuine Social Security Number paired with a fabricated name and date of birth.

The reason this works in the United States is structural. SSNs were never designed as identity credentials: they were designed to track earnings for benefit calculation. There is no photograph, no biometric, and no authoritative public mechanism for confirming that a given SSN belongs to a given name.

The problem worsened in 2011 when the Social Security Administration randomized SSN issuance. Before randomization, the first digits encoded geographic and temporal information, which allowed rough validation, an SSN inconsistent with a claimed birth year or state was a red flag. Randomization eliminated that signal entirely.

Synthetic identities are frequently built around SSNs belonging to people unlikely to notice misuse: children, the elderly, the deceased, and the incarcerated. The identity is then cultivated over months or years. Small credit accounts establish a file. Payment history builds a score. Eventually the identity looks, to any automated system, like an established person with a modest financial footprint.

What makes this threat particularly difficult is that there is no victim to report it. A stolen identity generates complaints. A synthetic identity generates nothing, because the person whose SSN was used has no relationship with the accounts opened.

Deepfake and synthetic media fraud

The second threat targets the verification process itself. Generative models can now produce identity document images that reproduce layout, fonts, holographic overlays, and machine-readable zones with high fidelity. They can produce photographs of people who do not exist. And they can produce video of a face responding in real time to prompts.

FinCEN's alert described exactly this pattern: criminals using GenAI to create falsified documents, photographs, and videos to bypass customer identification and verification controls, then using the resulting accounts to funnel proceeds and perpetrate downstream fraud.

The economics have inverted. What once required a specialist and physical materials now requires a subscription and a prompt.

---

Why Liveness Checks Stopped Being Sufficient

The industry's response to document forgery was to add a live capture step: have the user photograph or record themselves, confirm the face matches the document, and confirm the face belongs to a live human rather than a photograph.

This worked well against what security researchers call presentation attacks: holding a printed photo, a screen, or a mask in front of the camera. Detection of presentation attacks is a mature discipline, and good systems are effective at it.

It works considerably less well against injection attacks, which bypass the camera entirely.

In an injection attack, the adversary does not present anything to a camera. They insert media directly into the capture stream, using virtual camera software, emulators, or modified clients. The verification system receives what appears to be a live video feed. There is no physical scene to analyze, because there was never a camera involved.

This is the vulnerability FinCEN's red flags gesture toward when they reference customers using third-party webcam plugins during live verification checks, or attempting to change communication methods mid-verification, or citing technical difficulties to avoid a live check.

The structural implication matters for vendor selection. A verification flow that runs entirely in a browser, receiving a video stream without cryptographic assurance of its origin, cannot fully distinguish a genuine camera from a synthetic feed. Defenses against injection require attestation of the capture environment (establishing, cryptographically, that the media originated from a real sensor on a real device) which in turn requires SDK-level integration rather than a web page.

---

FinCEN's Red Flags, Read Operationally

The November 2024 alert provides indicators that translate directly into control design.

Document-level indicators. A photograph that is internally inconsistent or shows visual signs of alteration. A photograph inconsistent with other identifying information, a date of birth suggesting a person substantially older or younger than the image. Multiple identity documents presented that are inconsistent with each other.

Verification-process indicators. Use of a third-party webcam plugin during a live check. Attempts to change communication method mid-process. Reluctance or refusal to complete a live verification step, particularly after a document has been submitted.

Technical indicators. Media flagged by deepfake detection software. Profile text flagged by GenAI-detection tools. Geographic or device data inconsistent with the identity documents presented, a passport from one country presented from a device consistently located elsewhere with no explanation.

Behavioral indicators. A newly opened account with a pattern of rapid transactions. High payment volumes to higher-risk payees such as gambling sites or digital asset exchanges. High volumes of chargebacks or rejected payments.

The pattern in this list is instructive. Only a minority of the indicators concern the document itself. Most concern context: how the submission was made, from where, on what device, and what happened afterward. That is a deliberate signal about where detection capability actually lives.

---

Building Layered Identity Assurance

No single control defeats these threats. Effective programs layer independent signals so that defeating one does not defeat the system.

Layer one: real document authentication

Not "we received a photograph", but actual authentication. This means validating document structure against known templates for the issuing authority, checking security features that are difficult to reproduce, verifying machine-readable zone checksums, confirming internal data consistency across the document, and detecting digital manipulation artifacts.

Where the document contains an NFC chip, as U.S. passports and an increasing number of other credentials do, reading that chip directly is the strongest available document control. The chip data is cryptographically signed by the issuing authority. A generated image, however convincing, has no chip to read.

Layer two: injection-resistant capture

If the verification step matters, the capture environment must be trustworthy. That means SDK-based capture with device attestation rather than an unauthenticated browser stream, detection of virtual camera and emulator signatures, and analysis of signal characteristics consistent with genuine sensor output.

Layer three: identity data corroboration

For U.S. persons, the Social Security Administration's electronic Consent Based SSN Verification service allows an organization, with the individual's written consent, to confirm whether a submitted SSN, name, and date of birth combination matches SSA records. This is one of the few authoritative checks available against synthetic identities, because a fabricated name attached to a genuine SSN will not match.

Death Master File screening addresses SSNs belonging to deceased individuals. Credit bureau data provides identity depth signals: how long the file has existed, whether the address history is coherent, whether the identity has the accumulated footprint a real person of that age would have.

A thin file on a claimed forty-five-year-old is a strong synthetic indicator.

Layer four: device and behavioral signals

Device fingerprinting, IP geolocation and its consistency with claimed residence, velocity checks across submissions, email and phone number tenure, and behavioral characteristics during form completion. Automated form-filling patterns differ measurably from human ones.

None of these is conclusive alone. In combination they produce a risk picture that a synthetic identity struggles to satisfy across every dimension simultaneously.

Layer five: post-onboarding monitoring

The behavioral red flags in FinCEN's alert appear after account opening. Rapid transaction patterns on new accounts, payments to higher-risk payees, and chargeback volumes are all detectable only through monitoring. An onboarding control with no monitoring behind it has a single point of failure.

---

What This Means for Funds, Real Estate, and Law Firms

There is a tendency in non-bank sectors to treat this as someone else's problem. That is a mistake, and for a specific reason: adversaries target the weakest available control.

A private fund accepting subscriptions from individual investors, a title company verifying the identity of a beneficial owner for a Real Estate Report, and a law firm onboarding a client for a transaction all perform identity verification with substantially less infrastructure than a bank. They are, by that measure, softer targets.

The stakes are also higher per event. A bank's synthetic identity exposure is typically credit loss on a modest facility. A fund's exposure is an unverifiable investor in the register, potential sanctions liability, and the reputational consequence of explaining to institutional LPs how it happened. A title company's exposure is a federal filing containing false beneficial ownership information.

Two practical points follow. First, the sophistication of the identity verification should scale with transaction value, not with institutional category. A $3 million property purchase warrants stronger verification than a $200 account opening, regardless of whether the verifying institution is a bank.

Second, verification should not be delegated implicitly. Firms frequently assume that someone else in the transaction chain verified identity, the bank, the fund administrator, the introducing broker. Assumptions of this kind are rarely documented and rarely correct.

---

The Detection Arms Race, Honestly Assessed

It is worth being candid about trajectory. Detection of AI-generated media works by identifying artifacts of the generation process. Generation models improve continuously, and each improvement erodes the artifacts detectors rely on. This is a race in which the defender is structurally behind, because the attacker can test against available detectors before deploying.

The strategic conclusion is that controls which depend on distinguishing real media from generated media should not be load-bearing. The durable controls are the ones that do not require making that distinction:

Cryptographic verification, such as reading a signed NFC chip, where authenticity is proven rather than assessed.

Attestation of the capture environment, where trust derives from the device rather than from analysis of the content.

Corroboration against authoritative external records, where a fabricated identity fails because the underlying data does not exist.

Cross-signal consistency, where an adversary must simultaneously satisfy document checks, device signals, behavioral patterns, and external data, a substantially harder problem than producing one convincing artifact.

This is the architecture VeriKYC is built around: not a single verification step that must be perfect, but layered assurance in which each signal is independently recorded and the combination produces a defensible conclusion. Equally important, every signal and every decision is retained, so that if a fraudulent identity is later discovered, the firm can demonstrate exactly what it checked and what the checks returned.

---

Conclusion: Verify the Person, Not the Picture

The uncomfortable summary is that a photograph of an identity document is no longer evidence of anything. It is a claim, submitted by an unknown party, produced by unknown means.

That does not make identity verification impossible. It makes it a different discipline than it was five years ago. The question is no longer "does this document look genuine?" It is "what independent evidence do I have that this specific human being exists, is who they claim to be, and is the one submitting this application?"

Answering that question requires authenticating the credential cryptographically where possible, trusting the capture environment rather than the captured image, corroborating against sources the applicant does not control, and looking at behavior over time rather than a single moment of submission.

Firms still running a document-photo-and-selfie process and calling it verification are operating a control that FinCEN has publicly documented as defeated. The alert was issued in November 2024. The generation tools have improved considerably since. The reasonable planning assumption is that any control which merely inspects submitted media will continue to degrade, and that the controls worth investing in are the ones that never depended on inspection at all.